feat: multi-file upload with parallel transcription

[shaulams]

Summary

• Multi-file upload: Upload multiple interview recordings at once (different speakers/interviews) into a single project
• Parallel transcription: Files are transcribed simultaneously via ThreadPoolExecutor (up to 4 at a time) for faster processing
• Unified transcript: All transcripts merge into one view with time offsets, so clip selection and assembly work across all files
• Backward compatible: Old projects with single source_file auto-migrate to source_files array on load

How it works

1. Upload zone now accepts multiple files (<input multiple>)
2. Backend saves all files, spawns parallel Whisper transcription threads
3. After all complete, transcripts merge with cumulative time offsets (file B starts where file A ends)
4. Each segment/word gets a source_index pointing to its source file
5. Clip cutting resolves the correct source file and real timestamps via resolve_source_for_clip()

Test plan

• Upload a single file — verify transcription works exactly as before
• Upload 2+ files — verify all get transcribed and transcript merges correctly
• Mark clips across file boundaries — verify clip cutting produces correct audio
• Load an old saved project — verify it loads without errors (backward compat)
• Test with files that need compression (>25MB) — verify dynamic bitrate works

🤖 Generated with Claude Code